How much does an AI agent cost for a small business in 2026?

What are the five ways AI agents are priced?

Almost every quote you will get for an AI agent in 2026 uses one of five billing units, and the unit matters more than the headline number. Per seat charges for each human who uses the console, which is cheap for a small team but says nothing about agent volume. Per outcome (also called per resolution) bills only when the agent closes a conversation without a person, so you pay for success and nothing for escalations. Per conversation bills every conversation the agent touches whether or not it resolved it. Credits or actions meter the individual steps an agent takes (a lookup, a message, a tool call), which suits internal automations but makes customer-facing cost hard to predict. Per minute is the standard for voice agents and usually stacks the platform fee, the language model, text-to-speech and telephony.

The table lists one published example for each model. HubSpot's figure is reported by third parties rather than taken from a HubSpot price page, so treat it as reported.

What does an AI agent cost per 1,000 customer conversations?

To compare the models like for like, fix a volume and a resolution rate and do the arithmetic. The table uses 1,000 inbound customer conversations a month and two resolution rates: 40 percent, which is close to the 38 percent average that builts.ai measured across four small businesses in a 60-day Intercom Fin test, and 60 percent, which matches the 59 percent a HubSpot partner agency reported for a family-run manufacturer in its 9 October 2026 write-up. Per-conversation and per-minute pricing do not change with resolution rate because you pay whether or not the agent succeeded.

Credit pricing is the hardest to normalise because vendors meter steps, not conversations, and a conversation may consume anywhere from one credit to dozens. The Lindy row assumes five credits per conversation purely for illustration; replace it with the consumption you see in a two-week trial. Voice rows assume a three-minute average call.

Two things the table hides

First, the seat minimums are real: Intercom charges per human agent seat and HubSpot's outcome pricing sits behind a Service Hub Professional licence, so a three-person support team pays three seats before the first resolution. Second, outcome pricing rewards a good knowledge base. At 40 percent resolution Intercom Fin costs $0.99 per solved ticket and nothing for the 600 the agent hands over, while per-conversation pricing charges $2 for each of those 600 hand-offs too. If your help centre is thin, outcome pricing is cheap because the agent resolves little; it is also pointless for the same reason. Tidio's own blog lists its Lyro agent at a $29 per month Starter tier covering 100 billable conversations, according to Tidio, 2026, which is about $0.29 per conversation at that tier but needs a plan upgrade long before 1,000.

How much does a custom-built agent cost?

When no platform fits, or the agent has to read and write in your own systems, you build. Published small-business figures are narrower than you might expect. DevCom, 2 March 2026 puts a single custom workflow for a small or mid-sized business at $7,000 to $15,000. Adevs, 2026 frames first-year AI agent integration for small businesses at $12,000 to $180,000, with the top of the band reached only when several systems, compliance work and ongoing engineering are involved. The spread between the two is the difference between one well-scoped workflow and a programme.

The hours-times-rate framing explains both numbers. SecondTalent, 24 September 2026 lists US freelance AI agent developers at $95 to $142 an hour for junior, $119 to $185 for mid-level and $155 to $235 for senior, agencies at $180 to $490 an hour, and offshore developers at $25 to $35 an hour. Arc.dev's 2026 rate guide puts AI and machine learning engineering at $110 to $190 an hour. At mid-level US freelance rates, DevCom's $7,000 to $15,000 corresponds to roughly 40 to 125 hours of work, which is a realistic envelope for one workflow with two or three integrations, an evaluation set and a human-review step. The same hours at offshore rates come to roughly $1,000 to $4,400, and at agency rates to roughly $7,200 to $61,000, which is why three quotes for the same brief can differ by five times or more.

What pushes a build above these bands is autonomy (how many decisions the agent makes alone), the number of systems it must act in, the accuracy bar, data readiness and regulation, not the choice of model. Those drivers, the per-complexity market bands and the team shape are covered in detail in our guide to AI agent development cost, so they are not repeated here.

What are the running costs after launch?

A build quote is the start of the bill, not the whole of it, and this is where small businesses most often get surprised. Four recurring lines matter.

• Model and API usage. Every provider meters per token, and agents are token-hungry because each step re-reads instructions, context and tool results. Prices differ by model tier and change often, so take the number from a two-week pilot on real traffic rather than a demo. Caching repeated context and choosing the smallest model that passes your tests are the two biggest levers.
• Maintenance. An engineering write-up on dev.to, 2026 puts post-launch maintenance at 15 to 25 percent of the original build cost per year, covering prompt and tool fixes, integration changes and monitoring. On a $10,000 build that is $1,500 to $2,500 a year, which is comparable to a mid-tier SaaS subscription.
• Evaluation. Someone has to keep a test set of real cases, run it on every change and read the failures. On a platform this is your time; on a custom build it is a few engineer hours a month plus the model cost of the runs.
• Prompt drift after model deprecations. Providers retire model versions on fixed schedules, and an agent tuned to one version behaves differently on the next; a 19 April 2026 analysis of the AI feature maintenance cliff describes this as the point where unplanned rework lands all at once. Pin versions, keep evals and budget a re-tune on every forced migration.

The running cost of buying

Subscriptions bundle all of the above into the per-outcome or per-seat price, which is their real advantage at low volume: the vendor absorbs model deprecations and evaluation for you. What they do not absorb is the work of keeping your help centre and CRM clean enough for the agent to resolve anything. The HubSpot partner case above needed months of inbox and help-desk consolidation before its 59 percent rate was possible, and that owner time is the largest unbilled cost in most small-business deployments.

What does it cost for a 10-person company?

Two worked budgets make the trade-off concrete. Both assume a 10-person company with two people handling support escalations (so two seats where seats apply), a 40 percent resolution rate, five credits per conversation on the credits platform, model usage of $0.10 per conversation on the custom build (an assumption to be replaced with your pilot data), a $7,000 to $15,000 build at DevCom's band and maintenance at 15 to 25 percent of build cost in year one. All figures are month-one to month-twelve totals.

At 2,000 conversations a month the picture flips

Ten times the volume changes the ranking. Outcome pricing scales linearly with resolutions, per-conversation pricing scales with everything, and a custom build's large fixed cost is spread over 24,000 conversations.

Reading the two tables honestly

At 200 conversations a month nothing beats a subscription; the custom build costs five to ten times more per conversation and the only reason to choose it is a workflow no platform supports. At 2,000 a month a low-end custom build already matches Intercom Fin in year one, and in year two, when the build cost is gone and only maintenance and usage remain (roughly $3,450 to $6,150), it costs roughly a third to three-fifths of the subscription. If the resolution rate reaches 60 percent, Fin's year-one bill rises to about $14,952 and the crossover arrives sooner. The credits platform looks cheapest in both tables, and for an internal workflow it often is, but five credits per conversation is an assumption, credit consumption rises sharply with tool calls, and a credits product is not usually built for customer-facing support with escalation and audit trails. Run the same arithmetic with your own volume, your own measured resolution rate and real credit consumption from a trial before you decide; the full decision framework is in our guide to build versus buy AI agents for small business.

What is the cheapest way to get an AI agent?

The cheapest reliable route is to start no-code, scope one workflow with a measurable outcome, and use a free tier until the numbers justify paying. "An agent for customer service" is a programme; "an agent that answers the 20 questions that make up half our inbox and hands the rest to a person" is a two-week project that costs little to test. Free and entry tiers with their published limits as of October 2026:

• Zapier Free allows 100 tasks a month and Zapier Agents Free allows 400 activities a month, per Zapier pricing; enough to prototype a lead-reply or ticket-triage agent.
• n8n Community Edition is free to self-host, and n8n Cloud Starter is EUR 20 a month billed annually for 2,500 executions, per n8n pricing; self-hosting trades money for your own server and maintenance time.
• Intercom offers a 14-day trial, after which the Essential seat is $29 a month and Fin bills $0.99 per resolution, per Intercom pricing; at a few dozen resolutions a month the agent costs less than the seat.
• Make moved to credits on 27 August 2026 with Core at $10.59 a month for 10,000 credits, as reported by Jetadmin, 2026.
• Relevance AI's Pro plan is $19 a month billed annually, as reported by Layer3 Labs, 2026, and Microsoft Copilot Studio sells 25,000 Copilot Credits for $200 per pack per month, per Microsoft, October 2026, which only makes sense once you are already paying for Microsoft 365.
• Voice agents have no meaningful free tier, but Bland's Start plan charges $0.14 a minute with no platform fee, per Bland pricing, so a trial on 50 calls costs single-digit dollars.

What "cheapest" should mean

Measure cost per resolved conversation or per completed task against the baseline before the agent, not licence cost. A $29 plan that resolves nothing is more expensive than a $300 plan that closes 400 tickets. The practical sequence is: pick one workflow, write down today's cost in hours, run a free or trial tier for two weeks, count resolutions and escalations, then pay for whatever the data supports. Our guide to AI agents for small business lists the workflows that most often pass this test.

Are AI agents worth the cost for a small business?

On the survey evidence, yes for most businesses that get past the pilot stage, with an important gap between what vendors promise and what owners measure. Intuit QuickBooks' 2026 AI Impact Report, 12 May 2026, drawing on more than 34,000 survey responses and data from over 5.3 million small businesses, found that 78 percent of small businesses using AI say it improved productivity, up from 46 percent, and that of those paying for a dedicated AI tool in 2024, 86 percent were still paying in 2025. Goldman Sachs' 10,000 Small Businesses Voices survey of 1,256 owners, released 17 March 2026 and reported by Yahoo Finance, found 93 percent of users report a positive impact, while only 14 percent say AI is fully embedded in core operations.

The reality check is the resolution rate. The independent 60-day test by builts.ai across four small businesses averaged 38 percent of conversations resolved by Intercom Fin against marketing language of "up to 50 percent", and at the enterprise end Beam AI, citing Sinch's 2026 research, reports a 74 percent rollback or shutdown rate for agent deployments. Small businesses with a clean, narrow workflow tend to do better than that, as the 11-person edtech company Chalkie did in reaching 87 percent of support resolved without humans according to Intercom's own customer story, but that is a vendor-published best case, not a planning number.

Whether it is worth it for you comes down to four measurements, and every one of them is cheap to take during a trial: the share of conversations resolved without a person, the hours saved per week for the people who used to handle them, the change in response time and, where the agent touches sales, the change in conversion. Convert hours saved to the fully loaded cost of the people involved, subtract subscription or build-plus-maintenance cost, and you have a payback period you can defend.

How do you keep the cost down?

• Choose the billing unit that matches your risk: per outcome when your knowledge base is strong, per seat or credits for internal workflows, and avoid per-conversation pricing until the agent demonstrably resolves most of what it touches.
• Fix the knowledge base before the agent: every unanswered question is a billed escalation on per-conversation plans and a wasted opportunity on per-outcome ones.
• Scope one workflow and one channel, measure for two weeks, then widen; most failed deployments tried to automate the whole inbox on day one.
• Negotiate seat counts; only people who actually review agent output need a licence.
• On a custom build, use hosted model APIs, open-source frameworks and existing MCP servers rather than bespoke plumbing, pin model versions, and keep prompts, tool schemas and evals in your own repository so you are never locked to one engineer or vendor.
• Set per-request token and step budgets and cache repeated context so usage cost cannot run away on a busy day.
• Keep a human in the loop for write actions (refunds, bookings, record changes) until the logs prove the agent is reliable; a wrong action costs more than a slow one.
• Review the bill monthly against cost per resolved conversation; vendors change pricing models often (HubSpot and Make both did in 2026), and a plan that was cheapest in January may not be by August.

Where to get help building it

If the arithmetic above points to a custom build or a serious integration, the usual blocker is finding an engineer who has shipped agents with tools that can fail, not the budget. Hyparz matches pre-vetted remote AI agent engineers typically within days (about 5 percent of AI-role applicants pass vetting) and every engagement starts with a one-week risk-free trial, so you can scope the first workflow, see working code and keep whatever is delivered before committing. See our AI agent development services for how a build is scoped, and our comparison of AI automation agency versus in-house versus freelancer if you are still deciding which route to take.